基于神经网络的焊接缺陷智能化超声模式识别与诊断

INTELLIGENT ULTRASONIC PATTERN RECOGNITION AND DIAGNOSIS OF WELDING FLAWS BASED ON NEURAL NETWORK

以三种焊接缺陷为对象,研究了缺陷回波特征的评价与模式识别。在实验研究与理论分析的基础上,从每个缺陷回波样本中提取了26个特征值,采用基于统计学假设检验的特征评价和最佳特征子集选择方法,实现了特征空间的降维处理。作者采用B-P型反向传播神经元网络构成了智能化模式分类器,研究了网络模型的学习效果和对未知缺陷的分类识别能力。还探讨了用Dempster方法进行超声检测信息融合处理的可行性。实验结果表明,采用最佳特征子集作为样本的特征向量,获得了良好的识别结果,三类缺陷的平均正确识别率约为87.6％,最佳识别率为97％。

Experimental study and theoretical analysis on evaluation of flaw echo features and mode recogni-tion were carried out on the basis of three types of welding flaws. The dimentional reduction of feature space was brought out by the method of feature evaluation and optimum feature subset selection based on statistical hypothe-sis testing, in which 26 features were extracted from each echo samples. An intelligent pattern classifier with B-P type neural network was used in studying the learnt effect of network and the recognition ability for unknown flaws. At the same time, the feasibility of fusing ultrasonic information by means of Dempster was discussed. Experiments showed that the results of recognition was satisfactory when the optimum feature subset was taken as a sample's feature vector. The average recognition accuracy of the three types of flaws was about 87. 6%, and the best recognition accuracy amounted to 97%.